An informational approach to likelihood of malnutrition

Unidentified protein-energy malnutrition (PEM) is associated with comorbidities and increased hospital length of stay. We developed a model for identifying severe metabolic stress and likelihood of malnutrition using test patterns of albumin (ALB), cholesterol (CHOL), and total protein (TP) in 545 chemistry profiles. Pattern classes were derived by converting decision values to a number code using cutoff values for nonmalnourished (0), moderate (1), and severe (2) of: ALB 35, 27 g/L; TP 63, 53 g/L; and CHOL 3.9, 2.8 μmol/L. Patterns defined by combinations of normal and abnormal laboratory results had decreased the likelihood of PEM from an all-2 to all-0 pattern. They were compressed to four final classes. ALB (F = 170), CHOL (F = 21), and TP (F = 5.6) predicted PEM class (r2 = 0.806, F = 214; P < E−6), but pattern class was the best predictor (r2 = 0.900, F = 1200, P < E−10). Kruskal-Wallis analysis of class by ranks was significant for pattern class (E−18), ALB (E−18), CHOL (E−14), and TP (2E−16). The means and SEM for tests in three PEM classes (mild, moderate, severe) were: ALB-35.7, 0.8; 30.9, 0.5; 24.2, 0.5 g/L; CHOL-3.93, 0.26; 3.98, 0.16; 3.03, 0.18 μmol/L, and TP-68.8, 1.7; 60.0, 1.0; 50.6, 1.1 g/L. We classified patients at risk of malnutrition using truth table comprehension. The pattern classes formed by the tests are a better classifier than the tests themselves.